D^2-Monitor: Dynamic Safety Monitoring for Diffusion LLMs via Hesitation-Aware Routing

Aoxi Liu, Yupeng Chen, James Oldfield, Guanzhe Hong, Junchi Yu, Baoyuan Wu, Philip Torr, Adel Bibi

arXiv:2605.25893 · 2026-05-27 공개 · arXiv · PDF

diffusion-llms dynamic-routing safety-monitoring wildguardmix toxicchat openai-moderation bi-level-monitor efficiency-effectiveness

Abstract

Despite the emergence of diffusion large language models (D-LLMs) as an alternative to autoregressive large language models (AR-LLMs), safety monitoring for D-LLMs remains largely unexplored. Unlike AR-LLMs, D-LLMs generate text through a multi-step denoising process, exposing intermediate hidden representations that may contain safety-relevant information unavailable in standard single-step monitoring setups. Motivated by the suitability of lightweight probes for always-on monitoring, we analyze which trajectory-level signals best indicate when such probes are likely to struggle. We find that the most informative signal is safety hesitation: intermediate hidden states repeatedly falling within a small margin of the probe's decision boundary. The number of such hesitation steps in D-LLM's trajectory predicts probe failure effectively, providing a proxy of sample difficulty. Building on this analysis, we propose D^2-Monitor, a bi-level safety monitor for D-LLMs. D^2-Monitor adopts a lightweight probe as an always-on monitor to jointly estimate hesitation and perform base classification. When the hesitation level exceeds a threshold, a more expressive but computationally heavier probe is activated. This dynamic routing mechanism allocates monitoring resources efficiently at test time. Evaluated on 3 datasets (WildguardMix, ToxicChat, OpenAI-Moderation) across 4 D-LLMs, D^2-Monitor achieves state-of-the-art performance with a compact parameter footprint (leq 0.85M parameters), and exhibits the best trade-off between effectiveness and efficiency relative to 8 baselines.

한국어 요약

📋 한 줄 요약

**[Diffusion LLM 안전 / Dynamic Monitoring]** D^2-Monitor가 diffusion LLM의 trajectory hesitation 신호를 활용한 bi-level 안전 모니터로 ≤0.85M 파라미터의 lightweight probe로 SOTA 달성, 3 데이터셋·4 D-LLM에서 8 baseline 대비 best trade-off.

🎯 핵심 기여도

💡 핵심 아이디어

D-LLM의 multi-step denoising trajectory에 노출된 intermediate hidden state의 hesitation(probe boundary margin 반복 진입) 횟수가 sample difficulty의 강력한 proxy이며, 이를 활용한 bi-level routing이 모니터링 자원을 효율적으로 할당해 lightweight cost로 SOTA 성능을 달성한다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: D-LLM 안전 모니터링이라는 신생 영역의 첫 체계 연구, multi-step denoising의 hidden state 노출을 monitoring signal로 활용하는 새 패러다임, lightweight·dynamic routing으로 실용성·효율성 동시 확보. **한계**: D-LLM 4종에 한정한 평가, hesitation threshold tuning의 도메인 의존, autoregressive LLM에는 직접 적용 불가.

🚀 실용적 활용